Panorama generation faces a fundamental tension: equirectangular projection (ERP) enables global continuity but distorts polar regions, while cubemaps preserve perspective geometry yet suffer from seams. This paper proposes CubeAtlas, a compact single-canvas unfolding of cubemap faces with rotated half-polar regions, and pairs it with layout-aware sparse attention so a pretrained diffusion transformer can generate seamless 360° imagery in one forward pass without architectural surgery. The work matters because it promises both distortion-free perspective fidelity and global spherical coherence while retaining the strong priors of standard 2D image generators.

OmniRay tackles monocular 3D reconstruction across arbitrary central cameras without requiring intrinsics or camera-type labels at inference. Its core idea is that camera geometry has a hierarchy of complexity, so the ray field should be composed rather than represented by a single basis: a monotone radial base captures symmetric lenses, a spherical-harmonic residual handles smooth global deviations, and a free-form residual cleans up local variation, all applied as successive rotations on the sphere $S^2$. A DINOv2 backbone drives a Ray Module for $\mathbf{d}(u) \in S^2$ and a Distance Module for $\rho(u) > 0$, yielding an up-to-scale point cloud $\mathbf{P}(u) = \rho(u)\,\mathbf{d}(u)$ in one forward pass.

CODI tackles a key limitation in self-supervised Vision Transformers: the [CLS] token hogs global semantics onto a sparse set of representative patches, leaving other patch tokens weak for dense prediction. The authors show that image-level distillation gradients are gated by [CLS]-to-patch attention mass, which systematically under-trains non-representative regions. Their fix, Cross Distillation, mixes a small fraction of the teacher class-token distribution into each patch-token target, giving every masked patch an attention-independent semantic nudge. The method is simple, architecture-agnostic, and yields consistent gains on segmentation, depth, video object segmentation, and classification under aligned pretraining.

Joint audio-visual diffusion models repeatedly run both denoising branches and expensive cross-modal attention at every step, making inference costly. CAVC observes that audio-side drift caused by mid-stage caching can be repaired by a few terminal joint steps, whereas video-side degradation persists, and exploits this asymmetry with a training-free sandwich structure: full joint warm-up, video-only continuation with cached audio, and brief terminal recoupling. The result is a modest wall-clock speedup with improved visual dynamics, though the gains are bounded by the preserved video backbone.

Unsupervised federated classification with Vision-Language Models like CLIP suffers under non-IID shifts because standard pseudo-labeling discards ambiguous samples near decision boundaries. This paper proposes FedBST, a bilateral self-training framework that penalizes confident predictions on uncertain "waist" samples during training and refines inference by ensembling local visual prototypes with global textual prototypes. The goal is to squeeze extra robustness out of existing CLIP-based federated pipelines without any labeled data.

Episodic memory video question answering asks models to reason over long egocentric recordings by compressing the visual stream into a queryable memory. This work isolates retrieval within a fixed online textual-memory pipeline, systematically comparing retrieval units, context budgets, and reranking strategies on QAEgo4D-Closed. The authors show that their best Parent strategy matches state-of-the-art accuracy with far fewer tokens, and reveal that higher temporal retrieval coverage can actually hurt accuracy by injecting hard distractors.

This paper studies class-level unpaired vision-language matching, where frozen visual and language encoders must recover an unknown category correspondence without any paired supervision. Its core contribution is RankAlign, a preprocessing strategy that rank-transforms each within-modality similarity kernel before permutation matching, yielding an objective that is invariant to arbitrary strictly increasing reparameterizations of similarity values. The work matters because it tests whether the Platonic Representation Hypothesis can be operationalized through pure relational order rather than absolute metric alignment, and it provides exact-recovery guarantees under ordinal consistency.

Synthetic prostate MRI could help train deep learning models, but generating complete biparametric exams from anatomical masks remains underexplored. This paper proposes a 3D diffusion framework that compares sequential versus joint synthesis of T2-weighted and diffusion modalities, and evaluates directly synthesizing ADC and computed high-b-value DWI instead of deriving them from raw diffusion-weighted images. The work matters because it identifies practical pathways for generating clinically complete exams and shows that direct biomarker synthesis outperforms two-step derivation.

The paper tackles open-vocabulary semantic segmentation without pixel-level annotations, aiming to adapt CLIP from image-level classification to dense prediction. Its core idea is a regional semantic preservation module that alters the final attention layer to use a v-v self-attention mapping, together with a visual prototype memory bank and self-training with EMA to refine predictions. The prospect of annotation-free dense segmentation is appealing, but the paper’s claims and reproducibility need close scrutiny.

RaSFNet tackles LiDAR semantic segmentation by fusing 3D point features with 2D frustum features in a projection-based pipeline. Its three proposed modules—RAGDE for ray-aligned geometric encoding, STAM for azimuth-aware global context via a scan-serpentine Mamba, and RCPFF for range-constrained cross-representation fusion—are designed to resolve projection conflicts and preserve 3D spatial correspondence. The method reports $70.2\%$ mIoU on the SemanticKITTI validation set and $77.7\%$ on the nuScenes validation set.

Long chain-of-thought helps vision-language models solve complex visual tasks, yet it introduces a temporal grounding problem: intermediate reasoning steps can drift from image evidence, and standard final-answer reinforcement learning cannot credit which tokens should have revisited the image. This paper identifies "visual revisit"—localized spikes in cross-modal attention after language-internal reasoning—as a key temporal signal. The authors show across three backbones that revisit peaks predict correctness and are causally linked to accuracy. Building on this, they propose RECAP, a GRPO-compatible training method that propagates terminal rewards through revisit-conditioned dynamic discounting and gated value prediction, yielding consistent HallusionBench gains without inference-time cost.

GSHeadRelight tackles real-time, 3D-consistent portrait relighting from in-the-wild images by combining a generative 3D Gaussian Splatting backbone with a lightweight Unified Neural Radiance Transfer model. It assigns per-Gaussian spherical-harmonics coefficients that are view-conditioned to model low-frequency specular effects, trains on FFHQ under a white-light assumption without costly light-stage data, and achieves roughly 240 FPS by making relighting a simple per-Gaussian linear operation after a single forward pass.

Fetal congenital heart defect diagnosis is severely bottlenecked by the scarcity of positive cases, especially for rare subtypes that can occur in fewer than 1 in 10,000 births. This paper proposes Pos4CHD, a framework that synthesizes pseudo-positive echocardiography samples using only normal images and semantic masks, completely bypassing the need for real pathological training data. By combining clinically guided mask manipulation with a specialized generator called FE-Gen, the authors demonstrate that synthetically augmented models can approach or exceed real-data baselines on classification, detection, and segmentation tasks.

AgentVQA unifies 23 fragmented agentic datasets into 1,000 manually verified multiple-choice questions across five domains. By doing so, it provides a scalable offline complement to costly interactive simulator evaluations for vision-language models. The paper shows that even top models achieve only 58.5% accuracy and that failures stem primarily from grounded visual perception rather than reasoning mistakes.

Training-free personalization of vision-language models usually builds a database of isolated concept descriptions, which fails when user concepts are fine-grained and confusable. ReCoG argues that recognition is inherently relational: it constructs a Concept Graph whose directed edges store pairwise discriminative cues elicited from a frozen VLM, then retrieves by having candidates compete in explicit pairwise comparisons aggregated via round-robin voting. This yields large gains on fine-grained benchmarks and consistent improvements on standard ones, with strong ablations showing the edge content—not just the voting scheme—drives performance.

GraphCAD tackles generative modeling for CAD by replacing command sequences with Edge-Adjacency Graphs (EAGs), where wireframe edges are nodes and shared B-rep vertices define adjacency. It trains a hybrid graph diffusion model that jointly denoises discrete curve types and topology along with continuous geometric parameters, grounding the denoising network via an SE(3)-equivariant point-cloud backbone that renders latent edge parameters into sampled 3D curves. The work matters because sequence-based CAD generators propagate topological errors, while prior graph methods relied on complex face-centric surfaces that are hard to constrain.

Vision-Language Models typically encode an image in a single, task-agnostic pass before reasoning, which creates a bottleneck: the model can think longer but cannot refine the visual evidence itself. LoopLook addresses this by inserting lightweight recurrent loops into a middle-to-late window of the vision encoder, conditioning each loop on the question and adaptively deciding when to stop. The result is anencoder-side recurrent perception mechanism that improves accuracy on demanding perception benchmarks with minimal inference-time overhead.

This paper challenges the prevailing FID/CLIP-based evaluation of training-free diffusion samplers, showing that step-shortening methods collapse on complex generation benchmarks despite appearing lossless. The authors propose ParaLin++, a step-parity parallel integrator that preserves the full $N$-step discretization while accelerating solver iterations via a partially linear formulation with adaptive nonlinearization and successive over-relaxation. Experiments on FLUX and Ming-Flash-Omni 2.0 demonstrate that keeping the trajectory intact preserves semantic composition, dense instruction following, and long-text rendering where few-step solvers fail.

Intra-operative CT imaging of the liver often shows poor contrast for hepatocellular carcinoma, making tumour localisation difficult and curating labelled datasets expensive. This paper proposes guiding an RL-based search on intra-operative CT using weak supervision from a tumour-presence classifier trained on pre-operative MR images, eliminating the need for paired cross-modal data. The core idea is that an MR-trained classifier still assigns higher tumour probabilities to cancerous windows on CT, and these probabilities drive an agent to explore and aggregate a tumour score map. The work matters because it offers a way to transfer pre-operative appearance knowledge into the operating room without deformable registration or manual CT segmentations.

Multibeam imaging sonar is widely used for underwater perception, but its intensity images are degraded by coupled coherent speckle and system-induced blur, and clean reflectivity ground truth is almost never available. This paper derives the conditional-mean forward operator $A x = x * |h|^2$ from a stochastic coherent image-formation model and uses it to train a self-supervised reconstruction network from registered multi-frame observations. By projecting one frame's latent reconstruction through $A$ and matching it to a second registered frame, then regularizing with an adapted robust equivariant imaging loss, the framework targets joint despeckling and deblurring without any clean targets. The work matters because it replaces ad hoc intensity-domain speckle models with a physics-consistent inverse problem for real sonar platforms.